# 安装 Dockersudo dnf install -y docker
# 启动 Docker 并设置开机自启sudo systemctl enable --now docker
# 验证安装是否成功docker --version

# 将当前用户加入 docker 组sudo usermod -aG docker $USER
# 刷新组权限使配置立即生效newgrp docker

docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11-devel

# 查看 NPU 设备ls /dev/davinci*# 预期输出：/dev/davinci0  /dev/davinci2  /dev/davinci_manager ...
# 查看 NPU 状态npu-smi info

# 以 Qwen3-0.6B 为例ls /home/models/Qwen3-0.6B/# 预期包含：config.json  tokenizer.json  model*.safetensors ...

docker run -d     --name llama-cpp-ascend-910b     --device=/dev/davinci2     --device=/dev/davinci_manager     --device=/dev/devmm_svm     --device=/dev/hisi_hdc     -v /usr/local/dcmi:/usr/local/dcmi:ro     -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi:ro     -v /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro     -v /etc/ascend_install.info:/etc/ascend_install.info:ro     -v /home/models:/models     -v $(pwd)/llamacpp-workspace:/workspace/home     -v $(pwd)/pip-cache:/root/.cache/pip     -w /workspace     -p 8001:8001     --shm-size=16g     --ulimit memlock=-1     --ulimit stack=67108864     -e OMP_NUM_THREADS=8     swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11-devel     sleep infinity

docker exec -it llama-cpp-ascend-910b bash -l

# 容器内执行——先 source CANN 环境变量source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 替换 openEuler 国内加速 yum 源sed -i -e 's|https://repo.openeuler.org/|https://mirrors.huaweicloud.com/openeuler/|'        -e '/^metalink=/s/^/#/' /etc/yum.repos.d/openEuler.repo
# 安装编译依赖command -v cmake  >/dev/null 2>&1 || dnf install -y cmakecommand -v git    >/dev/null 2>&1 || dnf install -y gitcommand -v gcc    >/dev/null 2>&1 || dnf install -y gcc gcc-c++ makecommand -v ccache >/dev/null 2>&1 || dnf install -y ccache

git clone --depth 1     https://ghfast.top/https://github.com/ggml-org/llama.cpp.git     /workspace/home/llama.cppcd /workspace/home/llama.cpp

cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release       -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccachecmake --build build --config release -j $(nproc)

ls -lh build/bin/llama-serverbuild/bin/llama-server --version

ls /models/Qwen3-0.6B/*.gguf# 直接使用已有的 gguf 文件

pip install numpy sentencepiece transformers     -i https://mirrors.huaweicloud.com/repository/pypi/simplepython3 /workspace/home/llama.cpp/convert_hf_to_gguf.py /models/Qwen3-0.6B     --outfile /models/Qwen3-0.6B/Qwen3-0.6B-f16.gguf     --outtype f16

source /usr/local/Ascend/ascend-toolkit/set_env.sh
/workspace/home/llama.cpp/build/bin/llama-server     --model /models/unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf     --host 0.0.0.0     --port 8001     --n-gpu-layers 99     --split-mode none     --ctx-size 8192     --alias Qwen3.5-9B     > /workspace/logs/llama_server.log 2>&1 &

while ! curl -fsS http://127.0.0.1:8001/health 2>/dev/null; do    echo "等待服务就绪..."    sleep 5doneecho "服务已就绪！"curl -s http://127.0.0.1:8001/v1/models | python -m json.tool

curl -sS http://127.0.0.1:8001/v1/chat/completions     -H "Content-Type: application/json"     -d '{        "model": "Qwen3-0.6B",        "messages": [{"role": "user", "content": "请用一句话介绍华为昇腾。"}],        "max_tokens": 256,        "temperature": 0,        "chat_template_kwargs": {"enable_thinking": false}    }'

source /usr/local/Ascend/ascend-toolkit/set_env.sh
/workspace/home/llama.cpp/build/bin/llama-bench     -m /models/unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf     -n 512     -p 2048     -r 3     -ngl 99     -sm none     | tee /workspace/logs/llama_cpp_benchmark.txt

pkill -f llama-serverexitdocker stop llama-cpp-ascend-910bdocker rm llama-cpp-ascend-910b

sed -i -e 's|https://repo.openeuler.org/|https://mirrors.huaweicloud.com/openeuler/|'        -e '/^metalink=/s/^/#/' /etc/yum.repos.d/openEuler.repo